Pedestrian Detection Based On Road Surface Extraction In Complex Traffic Scenes

In the21st century, the traffic accident has become a serious socialproblem. As pedestrians are the major participants in the road traffic, theyare also the most direct victims in traffic accidents and their lives andproperty suffer huge losses. The advanced driver assistance system (ADAS),which aims at reduce the number of traffic accidents and protect the safetyof pedestrians, has been paid more and more attention all around the world.The pedestrian protection system (PPS) which is based on the computervision is an important part of ADAS and on-board pedestrian detection is themost critical technical problem in the PPS.Compared to the pedestrian detection on static images, onboardpedestrian detection is facing some new difficulties, such as high real-timedemand, wide range of illumination variation and so on. Most of theon-board pedestrian detection algorithms are based on feature extraction andmachine learning method, but these algorithms can’t achieve real-timerequirement and can’t detect small pedestrians effectively due to lacking ofdetails, complexity of selected feature and large amount of candidatewidows. After analyzing the existing pedestrian detection algorithms, wefind that the detection time and the number of false alarm windows can bereduced by narrowing down research area. According to this, we presentedan on-board pedestrian detection method which is based on the road surfaceextraction in this paper, and the main procedures as following: firstly, roadsurface is extracted in the illuminant invariant color space; secondly, regionsof interest (ROIs) are extracted based on the road surface; thirdly,pedestrians are detected with histogram of oriented gradient algorithm (HOG)in these extracted ROIs.Based on the extraction of ROIs, the number of candidate windows isdrastically reduced, so the detection time is reduced; and by resizing the ROIs to a fixed size, the detection rate of small pedestrian is also improved.Experiment results show that this method can remarkably reduce the falsealarm rate, improve the detection speed, and significantly improve the smallpedestrian detection rate.